[R] package MASS - MLE of negative binomial distributions

[Rik Verdonck]

Dear list,



I have a question about the exact estimate of the maximum likelihood for a negative binomial fit. I'm trying to approach this in two different ways: the first one is a fit using the glm.nb method, and the second one is a fit using the fitdistr function for each condition separately, where I add up all log likelihoods. These two methods do not yield the same values for the log likelihood of the fit. They do yield the same log likelihood if all data are one group (no summation), so I assume I'm doing something wrong when I sum up log likelihoods. Am I not "allowed" to do this?


Example code:
library(MASS)
x<-c(601,619,637,609,594,499,494,507,477,450,400,367,428,359,400,276,260,262,304,342,216,189,152,231,200,104,85,85,85,112)
groups<-as.factor(c(rep("dist1",5),rep("dist2",5),rep("dist3",5),rep("dist4",5),rep("dist5",5),rep("dist6",5)))

glm.nb(x~groups)$twologlik

logliks<-NULL
for(group in levels(groups))
{
	NBfit<-fitdistr(x[groups==group],"Negative Binomial")
	logliks<-c(logliks,NBfit$loglik)
	rm(NBfit)
}

sum(logliks)*2


Many thanks!
Rik